📄 Linguistic Feature Tagging for Automatic Classification of 27 Closely-Related Quechua Varieties

Summary

Most NLP tools treat Quechua as a single language, usually defaulting to Southern Quechua. Machine translation output, for example, can mix varieties into forms no community actually uses. This paper addresses that problem by building a dialect classifier for 27 Quechua varieties. The authors assembled a manually label-checked corpus of about 8.5M words from archives (AILLA, OLAC and others), web scraping and Bible editions. Bible verses were aligned with Spanish and English versions. Spanish object clitics, detected with spaCy, helped locate Quechua verbs showing polypersonal (subject–object) agreement. Hand-annotated lexical variants and these verbal agreement patterns were added as explicit features to a transformer classifier. The authors also compared several subword segmentation schemes. Neural models clearly beat statistical baselines and classify dialects with high accuracy. The linguistic features add little when data is plentiful but help more in low-resource and cross-domain conditions. The study also offers practical advice for other low-resource languages: lexical tagging appears to give the most benefit for the annotation effort required.

Key points

• The corpus contains 8,515,609 words across 27 ISO-coded varieties from several Quechua subfamilies. It combines spoken transcriptions, legal, educational and Bible texts. The No-Bible condition covers only 10 dialects.
• Texts are split into 250-word chunks with a 50-word overlap and divided 85/15 into training and evaluation sets, without stratification.
• Parallel verse-level alignment with Spanish and English used es_core_news_md and en_core_web_md. The annotation subset was limited to Matthew, Mark and John.
• The segmentation methods compared were BPE, PRPE, PRPE+BPE and unigram. With full data, weighted F1 ranges from about 0.94 to 0.99. In low-data settings, macro F1 falls to roughly 0.38–0.66, and no method wins consistently.
• Verb-agreement features help most when generalizing from little data, while lexical tagging helps most under domain shift. Code and data are available at github.com/clairepost/Quechua_Classifier.